A fault tolerance method is provided for a distributed stream processing system. The method includes: obtaining, by a working node, information of a failed node, and receiving retransmission data sent by an upstream node, where the retransmission data is originally sent by a source node to the failed node; determining, according to the information of the failed node, whether the failed node is related to the working node; and if the failed node is related to the working node, processing the received retransmission data, and sending the processed retransmission data to a downstream node. In the present application, retransmission data sent by an upstream node is received by a working node, and after it is determined that a failed node is related to the working node, the received retransmission data is processed, so that the retransmission data sent by the upstream node is processed only when the failed node is related to the working node.